RL fine-tune of Qwen3-32B on the e2egit-large task distribution (end-to-end git operations). Training via SkyRL/GRPO on Jupiter (GH200, 16 nodes). Best-reward exported checkpoint.

Training metrics (step 50)

reward/avg_raw_reward: 0.8828
reward/avg_pass_at_8: 0.891 (estimated)
Peak reward across run: 0.9238 @ step 41

Hyperparameters

Algorithm: RLOO-N (GRPO variant)
LR: 9e-6, epochs: 2, max_steps: 60
train_batch_size: 64, n_samples_per_prompt: 8
max_prompt_length: 32768, strategy: fsdp2
Dataset: exp_rpt_e2egit-large

laion/a2-rl-e2egit_large-50-32B

a2-rl-e2egit_large — step 50

Training metrics (step 50)

Hyperparameters

laion/
a2-rl-e2egit_large-50-32B